Improving network pavement performance management using machine learning

Highway networks play a significant role in people’s daily life. With limited funding and increasing demand, it is critical for transportation agencies to maintain the condition of the highway network cost-effectively. This necessitates the development of sound network pavement performance models and optimization approaches to select appropriate maintenance and rehabilitation (M&R) strategies for pavement performance management at the network-level. A key challenge to apply the current pavement performance models to network-level pavement management resides in the discrepancy between the data used for model development and the network-level pavement performance data. This dissertation tackles this issue by developing pavement performance models with variables that can be readily accessed from network pavement management systems (PMS). The developed models are demonstrated to generate accurate and reliable predictions and correctly capture the effects of different variables. This dissertation provides insight to improve pavement performance models by exploring the data imbalance of the network pavement performance data. The study identified two types of data imbalance: one resulting from the intrinsic pavement characteristics and the other stemming from pavement life expectancies. The existence of data imbalance negatively impacts the deterioration model's performance for pavement classes with limited data. The effect of model accuracy on the treatment benefit has also been investigated. Based on a reliability-based approach proposed in this study, it is discovered that with the increase of the model error, the treatment benefit decreases, while the extra cost needed to achieve the same benefit as the reference scenario increases. Considering the need for a cost-effective friction management approach and the potential of deep reinforcement learning to solve sequential decision problems, this dissertation develops a friction management framework based on the double deep Q-network (DDQN) reinforcement learning algorithm. Through a case study, it is demonstrated that with the DDQN model, better network friction performance can be achieved than the current practice of TxDOT, proving the effectiveness of the DDQN-based network pavement friction management framework.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

Doctoral

Grantor

The University of Texas at Austin

Year dc:date.issued

2023

Author dc:creator

• Xu, Hongbin, Ph. D.

Advisor dc:contributor.advisor

• Prozzi, Jorge Alberto

Committee members dc:contributor.committeemember

• Hersh, Matthew A.
• Hong, Feng
• Machemehl, Randy B.
• Claudel, Christian

• Machine learning
• Performance modeling
• Friction management
• Pavement
• Network

Language dc:language.iso

en

Identifier URI

https://doi.org/10.26153/tsw/62879

OAI identifier oai:identifier

oai:repositories.lib.utexas.edu:2152/135559